55 lines
1.7 KiB
Python
55 lines
1.7 KiB
Python
from http.server import HTTPServer, BaseHTTPRequestHandler
|
|
from os import chmod
|
|
from pathlib import Path
|
|
from socket import AF_UNIX
|
|
from socketserver import UnixStreamServer
|
|
from urllib.parse import unquote
|
|
|
|
from torch import float16
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
|
|
# https://stackoverflow.com/questions/21650370/setting-up-an-http-server-that-listens-over-a-file-socket
|
|
class UnixHTTPServer(UnixStreamServer):
|
|
def get_request(self):
|
|
request, client_address = super(UnixHTTPServer, self).get_request()
|
|
return (request, ["local", 0])
|
|
|
|
|
|
class textgenHandler(BaseHTTPRequestHandler):
|
|
def do_GET(self):
|
|
prompt = unquote(self.path[1:])
|
|
print('Prompt')
|
|
print(prompt)
|
|
|
|
if prompt == 'favicon.ico':
|
|
return
|
|
|
|
input = tokenizer.encode(prompt, return_tensors='pt').to('cuda')
|
|
output = tokenizer.decode(model.generate(
|
|
input, do_sample=True, max_length=500, top_p=0.9)[0])
|
|
print(output)
|
|
|
|
self.send_response(200)
|
|
self.send_header('Content-Type', 'text/plain')
|
|
self.send_header('Content-Length', str(len(output)))
|
|
self.end_headers()
|
|
self.wfile.write(output.encode('utf-8'))
|
|
|
|
|
|
# Load model
|
|
print('Loading model')
|
|
tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-neo-2.7B')
|
|
model = AutoModelForCausalLM.from_pretrained('EleutherAI/gpt-neo-2.7B',
|
|
torch_dtype=float16, low_cpu_mem_usage=True).to('cuda')
|
|
|
|
|
|
# Create and start server
|
|
print('Starting server')
|
|
path = '/srv/http/pages/textgen'
|
|
Path(path).unlink(missing_ok=True)
|
|
server = UnixHTTPServer(path, textgenHandler)
|
|
chmod(path, 666)
|
|
print('Server ready')
|
|
server.serve_forever()
|